Howdy, Stranger!

It looks like you're new here. If you want to get involved, click one of these buttons!

Supported by

Durbin Test - non-parametric mixed ANOVA

Hello

Can anyone help me in understanding how to interpret the Durbin Statistic table for a two way rmANOVA where using non-parametric analysis. What is the difference between the p and the pF significance values?

Thanks

• Hi HMStat,

To facilitate the discussion it would be great it you could show a screenshot. We are working on ANOVA help files for the next release, so this may come in handy.

Cheers,

E.J.

• Hi HMSTAT,

The p is the p-value for the chi square statistic, whereas the pF is the p-value for the F statistic in the table.

Cheers,

Johnny

• Hi,

Maybe I don't understand well the output but I was wondering several questions related with this topic:

Why is not possible to run...

an unbalanced non-parametric mixed ANOVA?

between post-hoc test?

interaction post-hoc test?

• Hi Vicente,

• The Durbin test is only suitable for balanced designs (see Wikipedia https://en.wikipedia.org/wiki/Durbin_test ). I am not entirely sure why this is, but I can sort this out next week if needed.
• As for the the post-hoc tests, the between tests should be in JASP already, and the next release (0.10) will have tests for interaction terms.

Kind regards,

Johnny

• Hi JohnnyB,

Thanks for the replay.

These are very good news for the next release, I'm so impatient...

• Hi Johnny

I got lost in time, I'm back. I was unclear on your last response in April 2019 on what the Chi-square statistic and the F statistic are telling us with this test. I'm trying to test a repeated measure (group 1 variable - 4 levels), and there is a second independent variable which is the between subjects (2 levels). Is it right to place the between subjects variable in the Optional Grouping Factor box, and if so, what is the pF telling us (I take it the p-value of the Chi-square is that there is a significant difference between at least one of the repeated measures levels). But for pF I get significance even if I enter the same data for both levels of the grouping factor/between subjects. Unfortunately there is no item in the JASP help manual for Durbin, only for Friedman. Below is the output:

Regards

• Hi HMSTAT,

The Chi squared is the original test statistic as proposed by Durbin, whereas the F statistics is a slight variation on it, introduced later. (see https://en.wikipedia.org/wiki/Durbin_test#Test_definition for the history). For completeness sake, I thought to include both statistics.

As for the grouping variable - this is an optional factor that can be specified to distinguish between different groups. By default, the observations are grouped based on participant ID (like in the Friedman test), but if the grouping variable is specified, the observations are grouped based on that (between subjects) factor. In terms of the PMCRplus package and the description on Wikipedia, this is the "blocking" variable.

So you will still get a significant result in your case, if the treatment (i.e., within subject levels) differ from each other. What the grouping variable does is relax the assumption that the blocks are independent, so instead of assuming that all participants are independent of each other, it assumes that the different groups are independent of each other.

Does that clarify things? Please let me know if not =)

Kind regards,

Johnny

• Thanks Johnny, crystal clear!

• Hello,

I have two variables. One (Size) is having three levels, and the other (Time) is seven levels. I am trying to run a 3 x7 within-subject repeated measure ANOVA (non-parametric) and looking at the Durbin test. The test suggests the main effect of Size and Time. However, unlike parametric rmanova, jasp is not providing any interaction effect for non-parametric rmanova. Can anyone help me out with this? I am currently using the JASP- 0.12.2 version.

Thankyou!

• Hi Sandeep,

Unfortunately the possibilities for the nonparametric rm anova are not as extensive as for the parametric rm anova. The Friedman and Kruskal Wallis tests are suitable for one-way designs, but when we get to two-way designs, it becomes a bit tricky.

One possibility is to convert your dependent variable to ranks (where you pool the observations from all groups, assign ranks, and then split back into groups again), and then perform a parametric rm anova on the ranked observations. This process is described in this very nice paper of Conover and Iman: https://www.jstor.org/stable/2683975?seq=1

Kind regards,

Johnny